As co-founder and acting CTO of Atrieon at PSL, I've pioneered a revolutionary agentic orchestration platform that transforms how organizations manage complex projects, workflows and teams of AI Agents. Built to manage Agents from the ground up, Atrieon leverages cutting-edge AI to autonomously coordinate tasks, resources, and teams with unprecedented efficiency.

• Architected and implemented a first-of-its-kind agentic framework capable of understanding, planning, and executing multi-stage projects with minimal human intervention
• Developed proprietary orchestration algorithms that dynamically adapt to changing project requirements and resource constraints in real-time
• Created an innovative agent collaboration system where specialized AI agents work together to solve complex problems while maintaining alignment with human objectives
• Designed a natural language interface that allows non-technical stakeholders to define, monitor, and modify project parameters through conversational interactions

My work on the PROSE team represents the culimation of my dream to advance the state of the art in artificial intelligence. Throughout the years I have been exposed to the intricacies of the research methodology alongside cutting edge artificial intelligence techniques, and have worked alongside some of the best researchers in the world.

• Selected to assess the capacity and ability of OpenAI’s GPT-3, Codex and ChatGPT. Developed benchmarks and evaluation techniques, appropriate domains of application and future areas of research.
• Prototyped several early-stage LLM-based developer tools in VsCode with a reciprocal research relationship with GitHub Copilot.
• Created and deployed a VsCode extension to extract key developer information to power a “repeated-edits” framework for code repair, which was deployed in universities across three countries.
• Built a containerised web application named “AITutor” to combine code repair technology, execution infrastructure, and real-time hints to scaffold learning in University of Michigan's Online Data Science Class. [1]
• Created three new Data Extraction experiences utilising the Program Synthesis SDK into Microsoft PowerQuery Desktop (.NET) and Web (ReactJs). These experiences are utilised by approximately 100,000 users daily. [2][3][4]

The COVIDWA.com project began 7 days after I'd moved to the United States to help people find appointments for COVID-19 vaccines. The senior citizens of Washington had been thrown into a free-for-all, expected to search thousands of websites individually for rapidly-closing appointments. We automated this process helping over 1,000,000 people directly and an additional 2,500,000 through our API. A few learnings from this project are called out below:

• There are limits to how fast you can scale an organisation. Eventually, an extra person leads to more overhead than impact.
• Despite your intentions being good, your success can contrast the failure of powerful people. The CovidWa project was built by volunteers, and scaled to support millions in weeks. Powerful companies had been paid $40+ million dollars with year timeframes and failed to deliver.
• The media loves a feel-good story.

Leveraging my extensive startup experience, I'm driving innovation within Microsoft's Azure Synapse team, focusing on enterprise-scale machine learning solutions that mirror the agility and performance demands of high-growth startups.

• Architected scalable ML pipeline solutions capable of handling petabyte-scale data processing with startup-like efficiency
• Led the development of new Synapse features that reduce time-to-insight for data scientists by over 40%
• Implemented distributed training frameworks that enable ML models to scale horizontally across thousands of nodes

The AI-Tutor project combined the power of modern Large-Language Models and Program Synthesis technology with verification via assignment unit-tests. Using a combination of code repair providers, the project repairs the student's code, finds an explaination for the repair, and then turns that explaination into a "hint" rather than a direct guideline in line with supporting pedagogy.

• Built using Containerised infrastructure, able to verify 100's of candidate fixes and return in seconds
• Easily integrated into VsCode and Juypter via simple client side API's
• Deployed in real-world courses in a number of top colleges in the United States

In an effort to reduce the load that emails put on a C-Suite employee, the email assistant reads and creates draft replies for all incoming emails for a GMail account. The user can then accept, edit, or decline to reply with any of the drafts. The user can also add configuration parameters such as "I'm very busy today" which will modify the draft emails accordingly.

Multi-modal program synthesis refers to the task of synthesizing programs (code) from their specification given in different forms, such as a combination of natural language and examples. Examples provide a precise but incomplete specification, and natural language provides an ambiguous but more "complete" task description.